7 CNN

Critique of MLP

Problem: MLP always reshape the whole pictures into a vector, so it does not respect the spatial structure of images Therefore, CNN is proposed to process spatial features

Convolutional Layer

We first define an image as $l* h$ pixels, where each pixel means a color vector

For each filter, we calculate a dot product between the pixel’s color vector with the corresponding one in the filter, producing $5*5$